Enhancing Transport Layer Capability in HAPS-Satellite
Integrated Architecture
C. E. Palazzi ([email protected])
Dipartimento di Scienze dell’Informazione, Università di Bologna, Via Mura Anteo
Zamboni 7, 40127 Bologna, Italy
Computer Science Department, University of California Los Angeles, Boelter Hall,
Los Angeles CA, 90095 USA
C. Roseti ([email protected]) and M. Luglio
([email protected])
Dipartimento di Ingegneria Elettronica, Università di Roma Tor Vergata, Via del
Politecnico 1, 00133 Rome, Italy
M. Gerla ([email protected]), M. Y. Sanadidi ([email protected])
and J. Stepanek ([email protected])
Computer Science Department, University of California Los Angeles, Boelter Hall,
Los Angeles CA, 90095 USA
Abstract. The use of HAPS/UAVs to enhance telecommunication capabilities has
been proposed as an effective solution to support hot spot communications in limited
areas. To ensure communication capabilities even in case of emergency (earthquake,
power blackout, chemical/nuclear disaster, terrorist attack), when terrestrial fixed
and mobile infrastructures are damaged or become unavailable, the access to satellites represents a reliable solution with worldwide coverage, even though it may
suffer from shadowing impairment, especially in an urban environment.
In this paper we approach an innovative and more challenging architecture foreseeing HAPS/UAV connected to the satellite in order to enlarge coverage and to
allow interconnection with very remote locations. In this scenario we have analyzed
TCP based applications proposing some innovative techniques, both at protocol and
at architectural level, to improve performance. In particular, we propose the use of
a PEP technique, namely splitting, to speed up window growth in spite of high
latency, combined with TCP Westwood as a very efficient algorithm particularly
suitable and well performing over satellite links.
Keywords: HAPS, UAV, TCP, Satellite, Splitting, TCP Westwood
Abbreviations:
ABSE – Adaptive Bandwidth Share Estimation
ARQ – Automatic Repeat Request
ERE – Eligible Rate Estimate
HAPS – High Altitude Platform Station
RTT – Round-Trip Time
TCP – Transmission Control Protocol
UAV – Unmanned Airborne Vehicles
c 2004 Kluwer Academic Publishers. Printed in the Netherlands.
°
wpc-haps05.tex; 29/10/2004; 12:08; p.1
2
1. Introduction
To ensure telecommunication capabilities in emergency scenario requires the use of challenging architectures. The concept of using unmanned objects [25, 2, 24, 12], flying or stationary (HAPS/UAV) at
relatively low altitudes has been previously introduced and proven effective for backup or capacity upgrade in high traffic areas (hot spots).
The satellite is intrinsically suitable to provide service in such a scenario, as when used with HAPS/UAV, the combined architecture is
capable of enhancing hot spot coverage with reasonable latency. While
the HAPS/UAV provide short-range wireless connectivity at high rates
even with small user terminals, the satellite can ensure large bandwidth for very long range connectivity, to reach remote headquarters/command posts all over the world.
In this paper we revisit the performance of the protocols developed
for Hot Spot scenarios when used in this novel, rather unconventional
environment. With the HAPS/UAV + Satellite connection we expect
much higher loss rates on the ground-to-UAV link than in conventional
Hot Spots because of distance, obstacles and UAV motion. To characterize such a scenario. classical channel models developed for satellite
environments can be suitably applied [19, 8]. The exploitation of path
diversity, both at theoretical and at simulation level, has also been
investigated. The corresponding channel models[18, 11, 21, 14, 13] are
suitable for a more complex scenario involving multiple HAPS/UAVs.
Moreover, the satellite link has a large propagation delay and may
introduce random loss of its own. Since much of the applications (e.g.,
web traffic, area maps, image files, emergency reports, etc.) will run
on TCP, it is important to evaluate the performance of TCP in this
environment. We propose to study two different ways of maintaining
TCP connections:
1. End to end connections, from ground user to Internet server. We
will evaluate different TCP protocol choices including the legacy
TCP New Reno and the newly proposed TCP Westwood.
2. Proxy server on board of the UAV. The idea is to split the TCP
connection on the HAPS/UAV and thus reduce the problem into
two more easier sub-problems, i.e., a very lossy link with short propagation delay; and a more reliable (or at least, more predictable)
link with very large propagation delay. In this case, different TCP
“flavors” may be needed for the different link characteristics.
A system with multiple HAPS/UAVs in the sky to provide more
extensive coverage can additionally be identified. This latter scenario
wpc-haps05.tex; 29/10/2004; 12:08; p.2
3
is clearly much more complex as it also introduces the possibility of
HAPS/UAV to HAPS/UAV communications (possibly on a separate
radio channel). Moreover, users in an “urban canyon” can handoff to
another HAPS/UAV if the first HAPS/UAV goes beyond a building out
of sight. The handoff must be smooth in order to prevent disruption
of ongoing sessions (e.g. video conference among emergency teams). In
this paper we will limit ourselves to single HAPS/UAV scenario. The
results can be easily extended, however, to multiple HAPS/UAVs flying
platforms.
2. System Architecture
We consider an urban scenario with mobile users (pedestrians and cars)
connected by a very efficient communication infrastructure comprising
the cellular system as well as a Mesh Network with Hot Spots placed in
strategic locations (busy street crossings, tall buildings, parks, shopping
malls, airport lounges, etc.) to access multimedia services. In particular, data services (mainly, access to the Internet from mobile users),
if the trend is confirmed, will see a dramatic growth in the next few
years, owing to the introduction of new mobile services such as location
based resource discovery (e.g., nearest drugstore), navigation support,
web access, etc. In this environment, that is becoming increasingly
dependent on communications, in case of emergency situation when
power goes out and shortly thereafter all the Hot Spots and Cellular
Base Stations are shut off, communications come to a standstill until
power and telecommunication systems are restored. In addition, communications are most needed to control vehicular traffic and to allow
users to “navigate” their way (or alternatively, be remotely guided)
out of the traffic congestion. At the same time, repair crews, police
and medical teams need efficient communications to coordinate their
work. A similar emergency scenario in an urban area occurs when
there is a chemical or nuclear disaster caused by human error, plant
break down, act of war, or terrorist attack. Again in such situations
the communication infrastructure will have been impaired while the
need for efficient communications remains critical.
In the depicted scenario, to deploy in a short time a system composed of several HAPS/UAVs to establish an emergency telecommunication infrastructure is very cost effective and relatively easy. Due
to their unmanned nature, HAPS/UAVs can be deployed very rapidly
and flown even in environments potentially dangerous to humans, as
those polluted by chemical fumes or nuclear radiations. Assuming that
the onboard radio is the same as the ground Hot Spot radio, once in
wpc-haps05.tex; 29/10/2004; 12:08; p.3
4
place, the HAPS/UAV will act like a Hot Spot to the customers on the
ground permitting them to communicate among one another but also
to access the Internet, or a remote command post, via satellite.
The HAPS/UAVs cruise through the “urban canyons” acting as
repeaters and propagating the received signal to a wireless receiver
on the ground, exploiting paths that involve other HAPS/UAVs, or
the GEO satellite connected to a remote gateway located far away
from the disaster area [22]. In this way shadowing impairment may
be mitigated. The satellite antenna installation represents a significant weight and cost factor in the HAPS/UAV implementation. An
alternative realistic scenario involves a mobile ground satellite station,
for example a truck equipped with Wi-Fi and satellite transceivers,
as depicted in Figure 1. In this case, HAPS/UAVs can be relieved of
satellite transmitter overhead. HAPS/UAVs can thus be very simple,
requiring very little maintenance and rare upgrades, while the latest
technology innovation can be easily implemented in the “on-wheels”
satellite station.
Figure 1. Scenario
This solution greatly enhances “prompt response” in case of emergency. Both trucks and lightweight UAVs can be deployed literally
within minutes after the accident. UAVs can fly at elevations of a
few thousand feet, well below commercial airliners. Heavy HAPS (with
satellite gear) on the other hand would require much more elaborate
and time consuming launching and navigation procedures. In addition,
the higher altitude with respect to UAV implies slightly different per-
wpc-haps05.tex; 29/10/2004; 12:08; p.4
5
formance in terms of link budget and delay. On the negative side, there
is an additional link on the end-to-end path (the link from UAV to
satellite truck). Moreover, since the UAVs in the urban emergency environment fly much lower than the HAPS in the operational environment,
it may be necessary to deploy multiple trucks, or a “constellation” of
multiple UAVs to provide adequate UAV-to-Truck connectivity.
The eventual presence of several flying units can greatly enlarge the
area covered by the proposed architectures and the exploitation of path
diversity can improve availability in an urban environment. Indeed, this
two-level “satellite empowered” architecture includes the advantages of
very small mobile terminal technology, the ability of handling high data
rates, and the capability to establish very long-range connections.
In the following subsection, we include a more detailed description
of the HAPS/UAVs features.
2.1. High Altitude Platform Stations (HAPS)
The proposed solution for an emergency urban environment demands
highly efficient broadband and multimedia services. Ideally, the goal is
to build a wireless network able to efficiently cover a wide area with
low propagation delays and negligible multi-path fading. In this context, radio communications could be based on higher frequency bands
looking for larger bandwidths. Unfortunately, the use of this spectrum
of frequencies implies line-of-sight propagation between the base station
and each customer terminal since local obstructions cause problems.
The concept of High Altitude Platform Stations (HAPS) is not
new. What is quite new is the willingness to utilize such platforms
as a structure to locate a communication payload. The capability to
offer connectivity over large areas utilizing a device located at a few
kilometers (in near space), represents a trade off solution between very
distant satellites and terrestrial repeaters. In fact, HAPS do not require
complex launch, can be easily moved from one position to another, and
can be easily repaired and re- launched. Moreover, free space losses
and propagation delay are not critical at all. The flying objects usable for this scope can be classified as stationary (blimps) or moving,
both usually unmanned (even though someone has proposed manning
such vehicles). The balloons have the advantage that they appear fixed
with respect to an Earth observer but the mechanical stabilization is
extremely critical. On the other hand, moving platforms need to be
tracked from the ground even though they actually fly over limited
areas. The coverage area of a platform typically has a diameter up to
200 km subdivided in cells of 1-10 km.
wpc-haps05.tex; 29/10/2004; 12:08; p.5
6
HAPS located in the stratosphere at 17-22 km can carry communication relay payloads and operate in a quasi-stationary position. The
payload may include a complete base-station and a satellite transponder. Therefore, HAPS can combine the best features of both terrestrial
and satellite systems leading to a powerful integrated network. A single
HAPS can replace a large number of terrestrial resources covering a
wide area by providing a flexible cellular frequency re-use structure
with a reduction in service costs. Basically, HAPS are aircraft or airships (essentially balloons, termed “aerostats”) and can be manned or
unmanned; in the latter case, autonomous operations can be coupled
with remote control from a control ground station. Throughout the
evolution of HAPS, three types of vehicles have been distinguished:
1. Unmanned Airships. These can range from small expendable balloons flying at modest altitudes to large airships operating at about
21 km altitude. Typically, the balloons are solar-powered and filled
with helium.
2. Unmanned Aircrafts. This type of HAPS is an unmanned solarpowered plane, which can fly against the wind or in a roughly
circular tight path to maintain a position as stable as possible.
These platforms are power-limited and need to store sufficient energy for station- keeping throughout the night. A fuelled unmanned
version of these aircrafts, flying generally at modest altitudes, is
specifically known as Unmanned Aerial Vehicle (UAV).
3. Manned Aircraft. These are represented by the conventional fuelled
aircraft.
The advantages of HAPS communications can be summarized as:
Rapid deployment. Unlike satellites, the HAPS-based services can be
designed, implemented and deployed relatively quickly. This presents
obvious advantages in terms of rapid restoration of communication
resources, after a disaster or in emergency scenarios.
Relatively low cost upgrading of the platform. The possibility of
effortlessly landing and re-launching permits easy upgrades or reconfigurations of the payload, allowing a high degree of “future- proofing”.
Broadband capability. Potentially, a large number of users can be
served exploiting mm-wave frequencies with their large bandwidth allocations and offering line-of-sight links over many areas.
Flexibility to respond to traffic demands. HAPS can provide an efficient resource allocation through flexible and responsive frequency
re-use patterns and adaptable cell sizes.
wpc-haps05.tex; 29/10/2004; 12:08; p.6
7
Low propagation delay. Delays of platforms located in the stratosphere are negligible compared with satellite delay, thus providing remarkable advantages for Internet best effort (TCP) and interactive
applications.
3. Protocol Enhancement
For effective bandwidth access in the scenario presented above, efficient algorithms must be implemented to guarantee wireless error
resilience. In the following subsection we describe two complementary
schemes that have both already demonstrated significant improvements
in TCP performance in a wireless environment: TCP Westwood and the
splitting technique.
3.1. TCP Westwood
In TCP Westwood (TCPW) [20], the sender continuously monitors
ACKs from the receiver and computes its current Eligible Rate Estimate (ERE). This scheme relies on an adaptive estimation technique
applied to the ACK stream. The goal of ERE is to estimate the rate
a connection is eligible for depending on congestion and bandwidth on
the path, and thus achieving high utilization without starving other
connections. Research on active network estimation [9] reveals that
samples obtained using “packet pairs” often reflects physical capacity,
while samples obtained using “packet trains” gives short-term throughput estimates. Not having the luxury to estimate using active probing
packets, a TCPW sender carefully chooses sampling intervals and filtering techniques to estimate the eligible rate of a connection. Duplicate
acknowledgments (DUPACKs) and delayed ACKs are also properly
counted in ERE computation.
In all TCPW implementations, upon packet loss (indicated by 3
DUPACKs or a timeout), the sender sets the congestion window (cwnd)
and the slow start threshold (ssthresh) based on the current ERE.
TCPW uses the following algorithm to set cwnd and ssthresh.
wpc-haps05.tex; 29/10/2004; 12:08; p.7
8
if ( 3 DUPACKS are received )
ssthresh = (ERE * RTTmin) / seg_size;
if (cwnd > ssthresh) /*congestion avoidance*/
cwnd = ssthresh;
endif
endif
if ( coarse timeout expires )
cwnd = 1;
ssthresh =( ERE * RTTmin) / seg_size;
if ( ssthresh < 2 )
ssthresh = 2;
endif
endif
The Eligible Rate Estimates (ERE) are determined using a timevarying coefficient, exponentially-weighted moving average (EWMA)
filter, which has both adaptive gain and adaptive sampling. Let tk be
the time instant at which the kth ACK is received at the sender. Let sk
be the ERE sample, and ŝk the filtered estimate of the ERE at time tk .
Let (αk be the time-varying coefficient at tk . The “Adaptive Bandwidth
Share Estimation” (ABSE) version [27] of the filter is given by:
ŝk = αk ŝk−1 + (1 − αk )sk
(1)
k −∆tk
where αk = 2τ
2τk +∆tk , and τk is a filter parameter which determines
the filter gain, and varies over time adapting to round-trip time (RTT)
and other path conditions. Below, we assume a fixed τk and focus on
adaptive sampling.
P
dt >tk −Tk
d tj
j
In the filter formula, the ERE sample at time k is sk =
,
Tk
where dtj is the number of bytes that are reported delivered by the jth
ACK, and Tk is an interval over which the ERE sample is calculated.
To preserve both efficiency and fairness, TCPW exploits a continuously adaptive sampling interval T: the more severe the congestion,
the longer T should be. The time interval Tk associated with the kth
received ACK is appropriately chosen between two extremes, as depicted in Figure 2 depending on the network congestion level. The
sampling interval, in fact, ranges between Tmin and Tmax , where Tmin
corresponds to an ACK-pair inter-arrival time, while Tmax is set to
RTT.
To determine the network congestion level, the ABSE estimator
compares the “Achieved Rate” with the instantaneous sending rate
which is equal to cwnd/RT Tmin . A measure of the path congestion level
wpc-haps05.tex; 29/10/2004; 12:08; p.8
9
is thus obtained. The difference between the instantaneous sending rate
and the achieved rate, clearly reflects the bottleneck queue, thus revealing that the path is becoming congested. The larger the difference, the
more severe the congestion, and the larger the new value of Tk should
be.
When the kth ACK arrives, the estimator first checks the relation
between the latest ERE estimate ŝk−1 and the current cwnd value.
When ŝk−1 ∗ RTT min ≥ cwnd , indicating a path without congestion,
Tk is set to Tmin . Otherwise, Tk is set to:
Tk = RTT ∗
cwnd − (ŝk−1 ∗ RTT min )
cwnd
(2)
or upon rearrangement:
Tk = RTT ∗
³
cwnd
RTT min − ŝk−1
cwnd
RTT min
´
(3)
In ( 3), cwnd/RT Tmin is the expected sending rate, while ŝk−1 is the
estimated rate the network allowed. TCPW has been shown to provide
higher efficiency while remaining friendly to other flows coexisting with
a TCPW flow[26].
S ev ere C o n g estio n — T m a x
U pon ACK
R eceip t
A B S E C o m p u tes
C o n g estio n L ev el
S am p lin g T im e
In terv al T k
L in k U n d er U tilized — T m in
Figure 2. Illustration of sampling time interval Tk adapting to network congestion
level
3.2. Splitting
Another class of approaches involves additional or enhanced network
infrastructure. In this case, intermediaries perform processing on behalf
of TCP endpoints to the greater benefit of performance. This basic idea
generalizes to the so-called “Performance Enhancing Proxy” or PEP[5].
Of the many PEP schemes, one involves segmenting, or “splitting” the
TCP connection into segments. Splitting makes use of TCP gateways
that maintain multiple TCP connections with both other gateways and
end users. In fact, between gateways splitting may use a specialized or
optimised transport protocol [16].
wpc-haps05.tex; 29/10/2004; 12:08; p.9
10
Naturally, subdividing the connection results in reduced RTT for
packets transmitted in each subsection of the original path. But as
a consequence, the node hosting the gateway must include an implementation of TCP that forwards packets between connections. So the
reduction in response time comes at the expense of memory and processing. Moreover, splitting violates the end-to-end semantics of TCP,
that is, packets corrupted while in the gateway’s buffer cannot be recovered through TCP means and thus result in a connection failure. Having
noted these consequences, many of the same system-level arguments in
favour of splitting on-board satellites apply equally to other nodes along
the connection [23]. As another advantage, splitting can allow disparate
terminals to communicate through the gateway, that is, it supports
TCP terminals without requiring uniform end-to-end TCP. Some hops
may use multicast, and thus more effectively exploit the broadcast
nature of wireless transmission while still supporting TCP endpoints.
Splitting TCP gateways also provide more rapid recovery from errors
and improves the overall robustness of the end-to-end connection when
several links suffer from shadowing or high error-rates.
A competing approach to connection splitting introduces an ARQ
link layer. For example, the IEEE 802.11 MAC protocol used in wireless
LANs includes an ACK. This does achieve many of the previously
described advantages, but by operating at the link layer, an ARQ mechanism forces all flows to undergo ARQ, even those such as real-time
applications like Voice over IP for which ARQ is notably the wrong solution. Beyond this system design consideration, link-level ARQ can introduce delay variations that actually degrade TCP performance unless
TCP is further optimised to account for them.
Having considered the benefits of connection splitting, we now more
carefully consider the design of the TCP gateway. To effect connection
splitting, a TCP forwarding agent receives data and sends acknowledgements on incoming connections while sending and retransmitting
data and receiving acknowledgments on outgoing connections. When
the forwarding agent receives a valid incoming data packet, it acknowledges this packet and places the packet in memory to be sent
on the corresponding outgoing connection. Consistent with the TCP
sender, this packet remains in memory until acknowledged by the outgoing connection. Once this happens, the packet may be removed from
memory.
Note that an unrestrained sender can quickly overflow the agent’s
memory space. As a result, the forwarding agent must control the rate of
incoming senders. To accomplish this, the agent uses TCP’s advertised
window mechanism. By progressively limiting the size of the window
as memory fills, the agent also limits the rate of the sender. Thus,
wpc-haps05.tex; 29/10/2004; 12:08; p.10
11
as packets become backlogged on the gateway node, the advertised
window decreases, resulting in backpressure on the sender. By limiting
the advertised window to half of free memory allocated to that connection, the gateway eliminates the possibility of dropping packets due to
buffer overflow. However, when unacknowledged packets consume the
total allocated space, the agent will advertise an empty window, thus
stopping the sender. In this case, the agent must send an additional
acknowledgment when space becomes available in order to restart the
sender.
4. Simulation Scenario
Traditional TCP New Reno (TCPNR) [10] was used to baseline the
performance characteristics of our architecture. We subsequently compared these results with results obtained using TCPW, a transport
protocol designed to handle wireless errors. We used the NS-2 simulator
[6] to verify the performance gains obtained with by our proposed architecture. NS-2 is widely used for simulating networks and is especially
valued for it’s models of transport protocols.
4.1. The Simulation Platform
NS-2 provides substantial support for discrete event driven simulations
at various network levels. In fact, a network configuration can be simulated integrating physical, routing, MAC, transport and application
layers both on wired and wireless environment (LAN, satellite, etc.)
Distributed with source code, NS-2 allows adding new models and
functionality for the purpose of investigating different scenarios and
new protocols. In fact, to run our tests we enhanced version 2.1b8a of
NS-2 by adding some new modules written in C++. In particular, we
added code to simulate the behavior of TCPW and the splitting scheme
of th HAPS/UAV. Using this tool, we performed an exhaustive set of
experiments with the following parameters:
wpc-haps05.tex; 29/10/2004; 12:08; p.11
12
− Packet size
− Queue size at nodes and cache size at the proxy (when present)
− Location of the proxy (when present)
− Propagation delay of each link
− Capacity of each link
− Error rate on each link
− Number of flows present at on the channel
− Protocols involved
4.2. Simulation Configurations
Figure 3 depicts the general topology adopted in our simulation corresponding to the architecture of Figure 1. Here, each of the circles in
the picture corresponds to one of the actors in the proposed scenario,
more specifically:
− W: generic wireless device in the urban area
− U: HAPS/UAV acting as a signal repeater
− T: Truck with a satellite station (eventually with proxy on board)
− S: GEO satellite (eventually with proxy on board)
− G: gateway on the ground
The edges between the various nodes represent wireless connections.
Note that the propagation delay between W and T, through U, is
small considering the flying altitude of the HAPS/UAV. And the delay
between a both between T and S and between S and G is a GEO satellite delay (125ms). Previous work has characterized errors on wireless
links [7, 3]. For our scenarios, we used a uniform PER (Packet Error
Rate) of 0.1% between T and G [15, 4, 17]. For links between W and
T, considered a wide variety of urban conditions. Shadowing was not
considered because we assume the user and UAV always remains within
line-of-sight.
The bandwidth available on each link is 1 Mbps and the packetsize is 1500 bytes, thus the end-to-end connection has a pipe capacity
of about 42 packets. Each adjacent node has a buffer of 50 packets,
while the proxy has a buffer of 200 packets, if not otherwise specified.
Each simulation utilized different combinations of transport protocols,
wpc-haps05.tex; 29/10/2004; 12:08; p.12
13
Figure 3. Simulation configuration
various PERs on the ground-UAV-ground wireless links, presence or
absence of a proxy on the ground satellite station and on the satellite,
diverse dimensioning of the cache in the proxy and alternate direction
of the data flow. To summarize, the various alternatives include the
following:
− transport protocol:
•
TCPNR, TCPW
− PER on the link between W and U:
•
0.1%, 0.5%, 1.0%
− proxy on board:
•
•
on the ground satellite station: split enabled, split disabled
on satellite: split enabled, split disabled
− cache size on the proxy:
•
10, 20, 30, ., 240, 250 packets
− traffic direction
•
from W to G, from G to W
wpc-haps05.tex; 29/10/2004; 12:08; p.13
14
Using 20 independent simulation runs for each configuration we calculated the average throughput during a period of 230 seconds as well
as the time required to transmit a 5 Mbyte file.
5. Results
We will first consider performance when combining the use of TCPW
and a splitting proxy in the presence of different PERs on the link between the wireless device and the terrestrial satellite station T through
the airborne HAPS/UAV. Figure 4 shows the time required to transmit
a 5 Mbyte file from W to G.
200
180
Transfer Time (sec.)
160
140
120
100
80
60
40
20
0
0.001
0.005
0.01
PER (in the wireless link)
NewReno_End-to-End
Westwood_End-to-End
NewReno_Split (Proxy in T)
Westwood_Split (Proxy in T)
Figure 4. Time to transmit a 5 MByte file from W to G.
In this case, TCPW demonstrates a significant advantage over TCPNR.
In particular without connection splitting, TCPW requires only from
59.33% (with 0.1% PER) to 34.99% (with 1% PER) of the time required
by TCPNR. When an intermediate proxy is introduced at T, the gap
between TCPNR and TCPW performance decreases significantly. In
this case, TCPW only needs 72.52% (with 0.1% PER) to 64.70% (with
0.5% PER) of the time required by TCPNR. When applied separately,
TCPW or splitting both lead to a substantial performance improvement. TCPW generally outperforms traditional transport protocols
when employed on wireless links involving long satellite paths, however,
its transmission time is further reduced with splitting. Finally, splitting
shows constant performance for different PER, demonstrating its talent
in hiding wireless losses to the endpoints.
wpc-haps05.tex; 29/10/2004; 12:08; p.14
15
In Figure 5 we compare the average throughput achieved by different transport protocols, the eventual utilization of a splitting proxy
on node T and various PER on links connecting W and G. For each
configuration we ran 20 simulations averaging the number of bytes sent
in 230 seconds to calculate the average throughput achieved. Again,
TCPW coupled with a proxy shows the best results, with an average
throughput ranging from 923.53 kbit/s (with a PER of 1.0%) to 961.18
kbit/s (with a PER of 0.1%). It appear splitting hides the frequent
errors on the shortest wireless link from the rest of the connection:
for each transport protocol, the average throughput achieved remains
relatively constant regardless of PER.
1000
900
Avg Throughput (kbps)
800
700
600
500
400
300
200
100
0
0
0.002
0.004
0.006
0.008
0.01
PER (in the wireless link)
NewReno_End-to-End
Westwood_End-to-End
NewReno_Split (Proxy in T)
Westwood_Split (Proxy in T)
Figure 5. Average throughput over a 230 sec transmission from W to G.
For sake of completeness, a set of simulations was performed which
considered the reverse data flow from G to W and the results closely
follow the previous set. In this case, performance improves as a result
of TCPW dealing with wireless links and the splitting at T protects
the connection from the frequent errors present on the links between
W and T. The average times required to transmit a 5 MByte file from
G to W are shown in Figure 6, while Figure 7 illustrates the average
throughput attained on a 230 seconds simulation run.
The results confirm the performance advantage of using splitting.
Moreover, since proxies along the path generate ACKs faster than a
distant receiver (with ≈500ms RTT), the congestion window at the
sender increases faster, thus speeding up the transmission rate. To better illustrate how the presence of proxies along the path improve overall
performance, we present in Figure 8 the sending window of the last TCP
segment along the path. Specifically, we utilized TCPNR with a 1.0%
wpc-haps05.tex; 29/10/2004; 12:08; p.15
16
180
160
Transfer Time (sec.)
140
120
100
80
60
40
20
0
0.001
0.005
0.01
PER (in the wireless link)
NewReno_End-to-End
Westwood_End-to-End
NewReno_Split (Proxy in T)
Westwood_Split (Proxy in T)
Figure 6. Time to transmit a 5MByte file from G to W.
1000
900
Avg Throughput (kbps)
800
700
600
500
400
300
200
100
0
0
0.002
0.004
0.006
0.008
0.01
PER (in the wireless link)
NewReno_End-to-End
Westwood_End-to-End
NewReno_Split (Proxy in T)
Westwood_Split (Proxy in T)
Figure 7. Average throughput over a 230 sec transmission from G to W.
PER on the links between W and G. We then compared the pure endto-end case with a proxy at T and with two proxies placed at T and S. In
each this case, we computed the sending window of the TCP connection
terminated at G, thus observing the instantaneous throughput at the
receiver. With a pure end-to-end connection we naturally considered
TCP between W and G, while including one proxy we have taken into
account the connection between T and G and, finally, with two splits
we have measured the sending values between S and G.
wpc-haps05.tex; 29/10/2004; 12:08; p.16
17
Figure 8. Sending windows at the last sending/forwarding hop with no split, single
split at T, or double split at T and S.
Without splitting, errors heavily impact TCP performance. In this
case, the sending window increases for a small period before a wireless
loss halves its value. This behavior, continuously repeated, results in
a noticeable underutilization of the available bandwidth. Conversely, if
the connection is split at T into two parts, the high error-rate between
W and T remains hidden from the rest of the link, thus resulting in
higher sending rates for longer periods. Wireless losses, in fact, are
in this case rapidly recovered along the short link between W and T,
and T can store enough packets to efficiently utilize the long, but still
quite reliable, links between T and G. Finally, when the end-to-end
connection is split twice, the presence of the proxy on the satellite S
shortens the duration of the send-acknowledge feedback-loop between
T and G subdividing it into two cycles: the first one between T and S
and the second one connecting S to G.
Using shorter connections accelerates the sending window growth
and thus creates a pipelining phenomenon that increases overall throughput. Figure 9 compares split with end-to-end connections. In this scenario, a 5 MByte file is downloaded from W by G. In this case, TCPNR
reaps the most benefit from splitting by achieving about 91% of the
capacity in case of double split and thereby achieving performance
equivalent to TCPW.
wpc-haps05.tex; 29/10/2004; 12:08; p.17
18
200
Transfer Time (sec.)
180
169.404
160
Westwood
140
New Reno
120
100
80
68.176
62.718
60
45.529
43.726
43.908
40
20
0
No-Split
S-Split
D-Split
Figure 9. Avg time to transmit a 5MB file from W to G with no split, single split
at T, or double split at T and S.
We now consider optimizing the size of the buffer of a proxy at T.
Figures 10 and 11 consider a configuration with 1.0% PER between W
and T and compare the average throughput for TCPW and TCPNR
with and without splitting. Clearly, an small amount of packets stored
in the proxy substantially reduces throughput. All the configurations
analyzed in our simulations experience a steep performance degradation
if we set the cache size to a value lower than the double of the channel
capacity (84 packets). Under this threshold, in fact, since the advertised
window sent back by the proxy corresponds to half of the free memory
in cache, TCP wastes a considerable amount of available bandwidth.
Indeed, with an undersized cache capacity, the proxy quickly runs out
of data to transmit. The congestion window might be large enough
to permit further transmissions but the window goes unused since the
undersized cache does not allow any prefetching and the proxy has
exhausted the few packets stored.
In fact, a value twice the pipe size is a lower bound for the dimensions
of the cache while larger values clearly help with better utilization of
the available bandwidth. Utilizing the proxy’s buffer undeniably protects reliable links from being impacted by transient burst errors that
temporarily slow down packets traversing adjacent error-prone edges.
For completeness, we also report the average throughput achieved by
TCPW and TCPNR without splitting.
wpc-haps05.tex; 29/10/2004; 12:08; p.18
19
1000
Average Throughput (kbit/s)
900
800
700
600
500
400
300
200
100
0
0
50
100
150
200
250
Cache Size (Packets)
WestwoodABSE_Split (with proxy)
NewReno_Split (with proxy)
WestwoodABSE_End-to-End
NewReno_End-to-End
Figure 10. Performance achieved per proxy cache size. Transmissions from W to G,
with a single proxy on T.
1000
Average Throughput (kbit/s)
900
800
700
600
500
400
300
200
100
0
0
50
100
150
200
250
Cache Size (Packets)
WestwoodABSE_Split (with proxy)
NewReno_Split (with proxy)
WestwoodABSE_End-to-End
NewReno_End-to-End
Figure 11. Performance achieved per proxy cache size. Transmissions from G to W,
with a single proxy on T.
6. Conclusions
Guaranteeing telecommunication services in emergency scenarios and
providing large bandwidth without reducing performance requires both
a well balanced architecture combined with efficient protocols. In this
paper, we proposed the combined use of HAPS/UAVs and satellites
as an innovative and challenging architecture able to meet the requirements when terrestrial infrastructures are unavailable. In addition, we
evaluated use of TCP Westwood and TCP connection splitting at the
ground station and satellite link as a means of improving efficiency.
wpc-haps05.tex; 29/10/2004; 12:08; p.19
20
The simulation results confirm the utility of TCP Westwood in
error-prone environments. They also detail the benefits of connection
splitting. As expected, TCP New Reno benefits the most from splitting.
In fact, with two proxies, one on the ground and one on the satellite,
the performance of TCP New Reno equals that of TCP Westwood.
References
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
‘NS-2
Trace
Formats:
http://www.k-lug.org/~griswold/NS2/
ns2-trace-formats.html’.
Avagnina, D., F. Dovis, A. Ghiglione, and P. Mulassano: 2002, ‘Wireless
networks based on high-altitude platforms for the provision of integrated
navigation/communication services’. IEEE Communications Magazine 40(2),
119—125.
Balakrishnan, H., V. N. Padmanabhan, S. Sehan, and R. H. Katz: 1997, ‘Acomparison of mechanism for improving TCP performance over wireless links’.
IEEE/ACM Trans. Networking 5(6), 756—769.
Baldantoni, L., H. Lundqvist, and G. Karlsson: 2004, ‘Adaptive End-to-End
FEC for Improving TCP Performance over Wireless Links’. In: ICC 2004.
Border, J., M. Kojo, J. Griner, G. Montenegro, and Z. Shelby: 2001, ‘RFC
3135: Performance Enhancing Proxies Intended to Mitigate Link-Related
Degradations’.
Breslau, L., D. Estrin, K. Fall, S. Floyd, J. Heidemann, A. Helmy, P. Huang,
S. McCanne, K. Varadhan, Y. Xu, and H. Yu: 2000, ‘Advances in Network
Simulation’. IEEE Computer 33(5), 59—67. Expanded version available as
USC TR 99-702b at http://www.isi.edu/~johnh/PAPERS/Bajaj99a.html.
Càceres, R. and L. Iftode: 1994, ‘The Effects of Mobility on Reliable Transport Protocols’. In: 14th International Conference on Distributed Computing
Systems. pp. 12—20.
Corazza, E. and F. Vatalaro: 1994, ‘A Statistical Model for Land Mobile Satellite Channels and its Application to Nongeostationary Orbit Systems’. IEEE
Trans. on Vehic. Techn VT-43, 738—741.
Dovrolis, C., P. Ramanathan, and D. Moore: 2001, ‘What do packet dispersion
techniques measure?’. In: IEEE Infocom’01. Anchorage, Alaska.
Floyd, S. and T. Henderson: 1999, ‘RFC 2582: The NewReno Modification to
TCP’s Fast Recovery Algorithm’.
Fontán, F. P., M. A. Vázquez, S. Buonomo, E. Kubista, and A. Paraboni: 1998,
‘A Methodology for the Characterisation of Environmental Effects on Global
navigation Satellite System (GNSS) Propagation’. International Journal of
Satellite Communications 16, 1—22.
Karapantazis, S. and F.-N. Pavlidou: 2004, ‘Broadband from Heaven’. IEE
Communications Engineer 2(2).
Loreti, P. and M. Luglio: 2001a, 3rd Generation Mobile Communication Systems, UMTS and IMT-2000, Chapt. Satellite Diversity: a Technique to Improve
Link Performance and Availability for Multicoverage Constellations. Berlin:
Springer Verlag.
Loreti, P. and M. Luglio: 2001b, ‘A Generalized N-State Model to characterize
satellite diversity for arbitrary number of satellites in case of uncorrelated
channels’. IEEE Communications Letters 5(11), 447—449.
wpc-haps05.tex; 29/10/2004; 12:08; p.20
21
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
27.
Luby, M., L. Vicisano, J. Gemmell, L. Rizzo, M. Handley, and J. Crowcroft:
2002, ‘RFC 3452: Forward Error Correction (FEC) Building Block’.
Luglio, M., M. Y. Sanadidi, M. Gerla, and J. Stepanek: 2004, ‘On-Board Satellite ”Split TCP” Proxy’. IEEE Journal of Selected Areas in Communications
special issue on “Broadband IP Networks via Satellites” 22(4), 362–370.
Lundqvist, H. and G. Karlsson: 2004, ‘TCP with End-to-End Forward Error
Correction’. In: International Zurich Seminar on Communications (IZS 2004).
Lutz, E.: 1996, ‘A Markoff Model for Correlated Land Mobile Satellite
Channels’. International Journal of Satellite Communications 14, 333—339.
Lutz, E., D. Cygan, M. Dippold, F. Dolainsky, and W. Papke: 1991, ‘The Land
Mobile Satellite Channel—Recording, Statistics and Channel Model’. IEEE
Trans. on Vehic. Techn. VT-40.
Mascolo, S., C. Casetti, M. Gerla, M. Sanadidi, and R. Wang: 2001, ‘TCP Westwood: End-to-End Bandwidth Estimation for Efficient Transport over Wired
and Wireless Networks’. In: Proceedings of the Seventh Annual International
Conference on Mobile Computing and Networking (MOBICOM-01). New York,
pp. 287—297.
Mazzenga, F. and F. Vatalaro: 2000, ‘Channel Modeling and Performance
Evaluation of EO Systems’. In: AP2000 Davos. Switzerland.
Palazzi, C. E., C. Roseti, M. Luglio, M. Gerla, M. Y. Sanadidi, and J. Stepanek:
2004, ‘Satellite coverage in urban areas using Unmanned Airborne Vehicles
(UAVs)’. In: IEEE Vehicular Technology Conference. Milan, Italy.
Stepanek, J., A. Razdan, A. Nandan, M. Gerla, and M. Luglio: 2002, ‘The Use
of a Proxy on Board the Satellite to Improve TCP Performance’. In: IEEE
Global Telecommunications Conference (Globecom), Vol. 21. pp. 2957—2961.
Thornton, J., D. Grace, C. Spillard, T. Konefal, and T. Tozer: 2001, ‘Broadband Communications from a High Altitude Platforms’. IEE Electronics and
Communications Engineering Journal 13(3), 138—144.
Tozer, T. C. and D. Grace: 2001, ‘High-altitude platforms for wireless communications’. Electronics & Communication Engineering Journal 13(3),
127—137.
Wang, R., M. Valla, M. Sanadidi, B. K. F. Ng, and M. Gerla: 2002a, ‘Efficiency/Friendliness Tradeoffs in TCP Westwood’. In: Seventh IEEE Symposium
on Computers and Communications. Taormina, Italy.
Wang, R., M. Valla, M. Y. Sanadidi, and M. Gerla: 2002b, ‘Adaptive Bandwidth Share Estimation in TCP Westwood’. In: In Proc. IEEE Globecom 2002.
Taipei, Taiwan, R.O.C.
wpc-haps05.tex; 29/10/2004; 12:08; p.21
22
Author’s Vitae
Cesare Roseti graduated cum laude in 2003 in Electronic Engineering at University of Rome ”Tor Vergata”. In 2003 and 2004 he was
a visiting student at Computer Science Department of University of
California, Los Angeles (UCLA). Since 2004 he is a PhD student at the
Electronic Engineering Department and his research interests include
satellites communications and transport protocols in heterogeneous
(wired/wireless) systems.
Claudio Enrico Palazzi studied Computer Science at University
of Bologna, Campus of Cesena. He has been a student representative
in several bodies of University of Bologna and, in particular, from
2000 to 2001 he was part of the Board of Governors. In 2001 he received the Sigillum Magnum of Alma Mater Studiorum University of
Bologna. He graduated cum laude in 2002 with a Thesis on Transport Protocols in Wireless Environments. In 2003 he was the first
student enrolled in the Interlink joint PhD program in Computer Science by which he is currently a PhD student in Computer Science at
both University of Bologna and University of California, Los Angeles
(UCLA). His research interests include protocol design, implementation
and performance analysis for wired/wireless networks.
wpc-haps05.tex; 29/10/2004; 12:08; p.22
23
Michele Luglio received the Laurea degree in Electronic Engineering at University of Rome “Tor Vergata”.
He received the Ph.D. degree in telecommunications in 1994.
From August to December 1992 he worked, as visiting Staff Engineering at Microwave Technology and Systems Division of Comsat
Laboratories (Clarksburg, Maryland, USA).
He received the Young Scientist Award from ISSSE ’95.
Since October 1995 he is research and teaching assistant at University of Rome “Tor Vergata” where he works on designing satellite
systems for multimedia services both mobile and fixed, in the frame of
projects funded by EC, ESA and ASI.
He taught Signal Theory and collaborated in teaching Digital Signal
Processing and Elements of Telecommunications.
In 2001 and 2002 he was visiting Professor at the Computer Science
department of University of California Los Angeles (UCLA) to teach
Satellite Networks class.
Now he teaches Satellite Telecommunications and Signals and Transmission. He is member of IEEE.
Mario Gerla received a graduate degree in engineering from the
Politecnico di Milano in 1966, and the M.S. and Ph.D. degrees in engineering from UCLA in 1970 and 1973, respectively. After working for
Network Analysis Corporation from 1973 to 1976, he joined the Faculty
wpc-haps05.tex; 29/10/2004; 12:08; p.23
24
of the Computer Science Department at UCLA where he is now Professor. His research interests cover the performance evaluation, design and
control of distributed computer communication systems; high speed
computer networks; wireless LANs, and;ad hoc wireless networks. He
has worked on the design, implementation and testing of various wireless ad hoc network protocols (channel access, clustering, routing and
transport) within the DARPA WAMIS, GloMo projects. Currently he
is leading the ONR MINUTEMAN project at UCLA, and is designing
a robust, scalable wireless ad hoc network architecture for unmanned
intelligent agents in defense and homeland security scenarios. He is also
conducting research on QoS routing, multicasting protocols and TCP
transport for the Next Generation Internet (see www.cs.ucla.edu/NRL
for recent publications). He became IEEE Fellow in 2002.
Dr. M. Yahya “Medy” Sanadidi was born in Damanhour, Egypt.
He received his high school diploma from College Saint Marc, and his B.
Sc. in electrical engineering (computer and automatic control section)
from the University of Alexandria, Egypt. Dr. Sanadidi received his
Ph. D. in computer science from UCLA in 1982. He is currently a
research professor at the UCLA computer science department. As coprincipal investigator on NSF sponsored research, he is leading research
in modeling and evaluation of high performance Internet protocols.
He teaches undergraduate and graduate courses at UCLA on queuing
systems and computer networks. Dr. Sanadidi was a manager and senior
consulting engineer at Teradata/AT&T/NCR from 1991 to 1999 and
led several groups responsible for performance modeling and analysis,
operating systems, and parallel query optimization. From 1984 to 1991,
he held the position of Computer Scientist at Citicorp where he pursued
R&D projects in wireless metropolitan area data communications and
other networking technologies. In particular, between 1984 and 1987, he
lead the design and prototyping of a Wireless MAN for home banking
and credit card verification applications. From 1981 to 1983, Dr. Sana-
wpc-haps05.tex; 29/10/2004; 12:08; p.24
25
didi was an assistant professor at the computer science department,
University of Maryland, College Park, Md. There, he taught performance modeling, computer architecture and operating systems, and
was principal investigator for NSA sponsored research in global data
communications networks. Dr. Sanadidi has consulted for industrial
concerns, has co-authored conference as well as journal papers, and
holds two patents in performance modeling. He participated as reviewer
and as program committee member of professional conferences. His
current research interests are focused on congestion control and adaptive multimedia streaming protocols in heterogeneous (wired/wireless)
networks.
James Stepanek received his B.S. in Computer Science from Harvey Mudd College in 1994 and his M.S. in Computer Science from
University of California, Los Angeles, (UCLA) in 2001 where he is
currently enrolled in the Ph.D. program. He is also currently a Member
of the Technical Staff in the Computer Systems Research Department
of The Aerospace Corporation. His research interests include wireless
and satellite networks.
wpc-haps05.tex; 29/10/2004; 12:08; p.25
wpc-haps05.tex; 29/10/2004; 12:08; p.26